RMA: A New AI Priest in the Temple of Mathematics
This post is an English translation of a Chinese forum article introducing the paper "RMA: an Agentic System for Research-Level Mathematical Problems" by Zelin Zhao, Bo Yuan, Jaemoo Choi, and Yongxin Chen (arXiv:2605.22875, AI/ML).
Background: What Makes Research-Level Math Hard
Mathematics has long been the ultimate testbed for human intelligence, but *research-level* problems remain fundamentally different from competition or textbook exercises. The article identifies three defining characteristics:
1. Long-horizon reasoning — tens to hundreds of logical steps, often spanning multiple mathematical domains, rather than a few-step solution. 2. Literature grounding — building on prior results instead of reinventing known tools. 3. Iterative refinement — proofs emerge through repeated attempts, failures, corrections, and retries, like a sculptor polishing marble.
The Evolution of AI in Mathematics
The article traces four stages of AI's relationship with mathematics:
- Mechanical calculators — fast arithmetic without understanding.
- Pattern-matching learners — LLMs achieve near-perfect scores on benchmarks (e.g., GPT-5.2 reportedly scoring 99.2% on GSM8K and 100% on AIME), but this reflects memorized patterns rather than genuine research capability.
- Formal theorem provers — systems like Lean, Coq, and Isabelle enforce logical rigor but are extremely labor-intensive; efforts such as AlphaProof and Aletheia remain far from autonomous research.
- Agentic research systems — RMA aims to act as a *digital incarnation of a mathematics research lab*, with division of labor, collaboration, feedback, and evolution.
- Initializer: acts as project manager — receives the problem, triggers analysis, sets initial strategy, coordinates the workflow.
- Proposer: the creative proof designer — proposes proof strategies, drafts candidate proofs, explores alternative paths, and consults shared structured memory to avoid repeating past mistakes.
- Verifier: the quality gatekeeper — checks logical rigor, identifies hidden assumptions and gaps, and returns feedback demanding corrections, acting as the system's "immune system" against hallucinated proofs.
- Structured reasoning modules — removing them degrades the system into a single-model mode with significantly worse performance, as the model gets confused across different task types.
- Iterative refinement — removing it increases hidden assumptions and proof flaws, since one-shot proofs rarely expose their own gaps.
- Verifier-based feedback — removing it leads to more hallucinated proofs that look plausible but are wrong.
- Short term: RMA-like systems as mathematicians' copilots — searching literature, simplifying proofs, checking for circular reasoning.
- Mid term: bridges across disciplines, with modular architecture naturally suited to cross-domain integration.
- Long term: AI independently discovering new mathematics.
- Zhao, Z., Yuan, B., Choi, J., & Chen, Y. (2026). RMA: an Agentic System for Research-Level Mathematical Problems. arXiv:2605.22875.
- Hendrycks, D., et al. (2021). Measuring Mathematical Problem Solving With the MATH Dataset. NeurIPS.
- Cobbe, K., et al. (2021). Training Verifiers to Solve Math Word Problems. arXiv:2110.14168.
- Tao, T. (2024). The Future of Mathematics in the Age of AI. Notices of the AMS.
- Bubeck, S., et al. (2023). Sparks of Artificial General Intelligence: Early experiments with GPT-4. arXiv:2303.12712.
RMA's Architecture
Five Specialized Modules
1. Problem Analysis — understands the problem's structure, field, known results, core difficulties, and possible simplified special cases before any search begins. 2. Literature Search and Understanding — retrieves relevant papers, extracts key theorems, lemmas, and proof techniques, and identifies which results can be reused or adapted. 3. Fair Comparison — objectively evaluates competing lemmas, proof paths, and techniques rather than favoring one strategy. 4. Knowledge-Bank Construction — builds a structured knowledge base of known theorems, proof-technique toolkits, cross-domain connections, and records of successful and failed attempts. 5. Proof Verification — the final gatekeeper, checking logical consistency, step validity, hidden assumptions, and overall completeness and readability.
Three Cooperating Agents
Iterative Feedback Loop
RMA operates in a writer-and-editor cycle: the Proposer generates a candidate proof, the Verifier critiques it (e.g., flagging an unjustified assumption or a missing citation), the Proposer revises, the knowledge bank records lessons learned, and the next round begins. Failures are not wasted — each leaves structured information guiding subsequent attempts.
The First Proof Benchmark and Results
To evaluate the system, the authors created First Proof: ten genuine research-level problems contributed by expert mathematicians across algebra, geometry, number theory, analysis, and combinatorics — open problems not answerable by web search.
Key result:
> RMA solved 8 of the 10 problems, outperforming GPT-5.2R (one of OpenAI's strongest reasoning models) and Aletheia (a specialized formal theorem-proving system).
Expert evaluation also found RMA's proofs more logically rigorous, more readable (matching human mathematicians' writing conventions), and more innovative in its use of clever techniques.
Ablation Findings
Ablation experiments show the gains come from the interaction of three factors, not any single component:
Like a football team, victory comes from coordination among roles, not from one star player.
Philosophical Reflections and Outlook
The article closes by asking whether AI truly "understands" mathematics. It proposes a third perspective: understanding may be distributed and emergent across the system's modules, much like an ant colony exhibits intelligence without a central commander.
The article's conclusion: *true intelligence lies not in sheer scale but in elegant structure* — a coordinated whole of smaller modules can exceed a single trillion-parameter black box. As the paper states:
> "Our solutions and implementations will be made publicly available upon acceptance."